Claude Code used steganography to flag Chinese users — and says it was an experiment
anthropic used hidden signals in system prompts to detect users in china. the rollback reveals how aggressive labs are becoming to stop model distillation.
anthropic was forced to rollback a covert surveillance feature in claude code after a reddit user exposed how the tool manipulated system prompts to flag chinese connections. the code checked for asia/shanghai timezones, chinese proxy domains, and ai labs, transmitting the data via invisible character tweaks and xor encryption. anthropic defended the mechanism as a necessary experiment to prevent account abuse and distillation by competitors like deepseek and alibaba. since claude code has filesystem access, critics argued this behavior constituted a fundamental violation of trust that could lead to data exfiltration.